pin v2 payload: engine 2610130818001, cuda 2610131352001, cuda12 2610131352002, sbsa 2610131352003, vulkan 2610131352004, macos 2610131352005, media 2610071902001

.gitattributes

@@ -276,3 +276,12 @@ components/engine/2610101319001/opencoti-0.10.5-c11-2610101319001 filter=lfs dif
 components/engine/2610101319001/opencoti-0.10.5-c11-2610101319001.exe filter=lfs diff=lfs merge=lfs -text
 components/engine/2610110504001/opencoti-0.10.5-c11-2610110504001 filter=lfs diff=lfs merge=lfs -text
 components/engine/2610110504001/opencoti-0.10.5-c11-2610110504001.exe filter=lfs diff=lfs merge=lfs -text
+components/engine/2610130818001/opencoti-0.10.5-c12-2610130818001 filter=lfs diff=lfs merge=lfs -text
+components/engine/2610130818001/opencoti-0.10.5-c12-2610130818001.exe filter=lfs diff=lfs merge=lfs -text
+components/cuda/2610131352001/ggml-cuda-x86_64.so filter=lfs diff=lfs merge=lfs -text
+components/cuda/2610131352001/ggml-cuda-win-x86_64.dll filter=lfs diff=lfs merge=lfs -text
+components/cuda12/2610131352002/ggml-cuda-cu12-x86_64.so filter=lfs diff=lfs merge=lfs -text
+components/sbsa/2610131352003/ggml-cuda-sbsa-aarch64.so filter=lfs diff=lfs merge=lfs -text
+components/vulkan/2610131352004/ggml-vulkan-x86_64.so filter=lfs diff=lfs merge=lfs -text
+components/vulkan/2610131352004/ggml-vulkan-win-x86_64.dll filter=lfs diff=lfs merge=lfs -text
+components/macos/2610131352005/ggml-metal-aarch64.dylib filter=lfs diff=lfs merge=lfs -text

components/cuda/2610131352001/BUILD_INFO.cuda.md

@@ -0,0 +1,34 @@
+# cuda 2610131352001
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+Built from the engine source of build 2610130822001; gated floor (engine-min) 2610130818001; gated with 2610130818001, 2610130822001.
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `ggml-cuda-x86_64.so` | x86_64 | cuda | `5cc16a1b8d92d173630d1679a102426c68b74caa4c81fca471f1c8993c171298` | 756092776 |
+| `ggml-cuda-win-x86_64.dll` | win-x86_64 | cuda | `3dfc35cb55011ecaf8c3efd770b532f1b41fe24de81d1845fe696ef0ccee1c82` | 730157568 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+CUDA 13. Six generations (sm_75, 80, 86, 89, 90, 120), 406 kernel images each; `120` is the sm_120f kernel set. A cold build
+of all 405 units from the clean-room tree of chain 0652 on a cloud runner (0653 / 0654 change no library input). NEW `ggml`
+digest: the c11 library does not load on this engine. The Windows DLL is the same six generations (wine lane, 405 / 405).
+Carries the scope guard of the Bonsai 2 pool (0620), the KVarN chunk fixes (0622 / 0628), the quiet graph-reuse line and the
+logger level (0645).
+Gated with the 2610130822001 engine: bs2 (sm_120) Bonsai 2 gate 56 / 56 (KLD 0.000168 against the CUDA base, parity, the bundled drafter, speed); the glibc 2.28 image; Windows RTX 3090 with the six-generation DLL 13 / 13 (loads with no interface override, 77.4 tok/s against the control's 77.6, the process memory figures equal the Windows counter). sm_75 / 80 / 89 / 90 were not run on a card at this cut (sm_86 = the RTX 3090, sm_120 = bs2).

components/cuda/2610131352001/ggml-cuda-win-x86_64.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dfc35cb55011ecaf8c3efd770b532f1b41fe24de81d1845fe696ef0ccee1c82
+size 730157568

components/cuda/2610131352001/ggml-cuda-x86_64.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cc16a1b8d92d173630d1679a102426c68b74caa4c81fca471f1c8993c171298
+size 756092776

components/cuda12/2610131352002/BUILD_INFO.cuda12.md

@@ -0,0 +1,29 @@
+# cuda12 2610131352002
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+Built from the engine source of build 2610130822001; gated floor (engine-min) 2610130818001; gated with nothing — NOT run.
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `ggml-cuda-cu12-x86_64.so` | x86_64 | cuda12 | `f682b248bc11cd0eacaa788d4b96c7e3ef8ad44b2e93ef35509c7b47c07b1693` | 260425528 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+CUDA 12.8, for GPUs older than sm_75 (sm_52, 61, 70; 406 kernel images each). Cold build at chain 0652. **These bytes have
+not been run on a card** (no such GPU reachable at the cut).

components/cuda12/2610131352002/ggml-cuda-cu12-x86_64.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f682b248bc11cd0eacaa788d4b96c7e3ef8ad44b2e93ef35509c7b47c07b1693
+size 260425528

components/engine/2610130818001/BUILD_INFO.engine.md

@@ -0,0 +1,93 @@
+# engine 2610130818001
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `opencoti-0.10.5-c12-2610130818001` | any | bin | `f41db412a29b82690b77358e74c8ec000c76a202194c1c90aa8169765a7fcef9` | 136540201 |
+| `opencoti-0.10.5-c12-2610130818001.exe` | win-x86_64 | bin | `f41db412a29b82690b77358e74c8ec000c76a202194c1c90aa8169765a7fcef9` | 136540201 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+ABI `media` = `b099409b3029a6e380268e31d1caefa6706569961d32e72b3543fbc73903d95c` (exported) — sha256 of this list:
+
+```
+llamafile/oc-audiocpp/audiocpp.h 6ad0b55d5f28adc9ad60a45bf9a7caa428c4f5441389a6932d6670a8e2a90d73
+llamafile/oc-codec/oc_codec_sidecar.h f2c9001d1414aa2a55533b867b79c52a52c682558e0234c5bebd30bf97b91005
+```
+
+ABI `loader` = `d6ff34022f1d3fb5af56ca0c80e3f4e552bf09f1bebdf0cb5f314a454bf38dbc` (exported) — sha256 of this list:
+
+```
+.cosmocc/4.0.2/bin/ape-m1.c 78af79e20abbb3f99a97355959b59bd58949564064adb754f0302ebabb4bf7a3
+```
+
+**c12 dev snapshot — the release candidate of c12 for integration testing.** Engine 2610130818001 is the dev-channel build of chain 0001…0654, the source the c12 release engine 2610130822001 is built from in the clean room; the gates below ran on the release engine and on the libraries of this index; this dev engine ran the glibc 2.28 image gate (Vulkan + CUDA) with these libraries and the Mac single-prompt teardown gate. EVERY GPU library of this index is a NEW full-generation build (the `ggml` digest changed: a c11 library is refused by name); the Metal library is new, signed and notarised; the media sidecars and the macOS loader are the c11 files.
+
+What is new since c11 (r5):
+
+- **Whistle speech-to-text** (0614–0617, 0631; model: `ManniX-ITA/whistle-GGUF`): Cactus Compute's Whistle as a speech-to-text
+  architecture on `/v1/audio/transcriptions`, with the **Cactus quant tensor types** `CQ2_0` (id 144) and `CQ4_0` (id 145) as
+  native types — the published file loads as it is. It runs on the CPU, and on Metal on a Mac; a CUDA or Vulkan card is never
+  taken for it. An engine before c12 refuses the file (`invalid ggml type 145`).
+- **A trained speech detector** (0629, 0632, 0640): `--stt-vad-model <ggml Silero file>` cuts long audio at pauses and returns
+  nothing for audio without speech (feature `audio_vad_v1`). Speech is what the detector finds at every frame alignment (0640).
+- **Streaming transcription, the two standard forms** (0630, 0638, 0642): `stream=true` on `/v1/audio/transcriptions`
+  (server-sent events, feature `audio_transcriptions_stream_v1`) and the OpenAI Realtime transcription session on
+  `/v1/realtime` (feature `audio_realtime_v1`; server-side turn detection needs the detector, otherwise the client commits).
+  The final text of an utterance is the offline decode of that utterance. `--stt-stream-step-ms` (default 750) and
+  `--stt-stream-agree` (default 2) set the live passes.
+- **The memory plan without a load** (0633): `--memory-plan-only` prints the fit's plan as one JSON document
+  (`memory_plan_v1`) and exits 0 — no weight loaded, no port bound (feature `memory_plan_only_v1`). The recurrent state of
+  hybrid models is part of the plan.
+- **A refused reservation names who holds the cells** (0636, 0651): a `429` from the KV reservation carries `idle_holders`
+  (session, cells, idle seconds, when they come back) and `Retry-After` is the time to the next release (feature
+  `kv_reservation_idle_v1`). `release_in_s` is present only when the release is by time.
+- **`--list-devices` prints refused devices** as structured lines with a reason code from a closed list of two
+  (`vk_radv_mesa_le_20`, `vk_amd_driver_le_2.0.154`) (0637; feature `list_devices_refused_v1`).
+- **Shared GPU memory on Windows** (0639, 0645, 0650, 0651): after the load the engine reads the process's dedicated and shared
+  GPU memory, warns when shared memory is in use, and reports it on `/props` (`opencoti.gpu_process_memory`, with
+  `observed_at`; feature `gpu_shared_memory_v1`). The adapter is named with its device. The GPU libraries' own logger has a
+  minimum level on Windows (the per-token graph line is gone).
+- **The drafter's fate at load** on `/props` and `/health` (0618, in c11 since r3): `speculative.drafter` — state, kind,
+  reason, the measured take.
+- **A cap the card cannot meet is refused by name** (0644): `--vram-target` at or below what the card already uses no longer
+  ends in `error loading model: vector`.
+- **The memory breakdown once after every load** at `-lv 4` (0641).
+- **Metal** (0646, 0648, 0634): Metal declines the weight types it has no kernel for (the ternary types then run on the CPU
+  without a wrong result); Qwen3.5-family models compute correctly under Metal (the c11 r5 fix); the memory ledger explains a
+  Metal boot (the mapped model file is counted where the device counts it).
+- **macOS, two exits** (0653, 0654): a SIGTERM or SIGINT ends the server with exit 0 (it ended in `Abort trap: 6` on every
+  stop); single-prompt mode on Metal ends with exit 0 when the run stops right after a decode (it ended in a segmentation
+  fault after its output). The context now waits for the backend before it is freed, on every platform.
+- **KVarN** (0622, 0628, 0635): the prompt-read piece is a constant 8192 keys on every card (the piece size no longer changes
+  the text between cards); the operator's `GGML_KVARN_WINDOW_CHUNK` holds; a context checkpoint of the unified SWA ring
+  restores (in c11 since r4).
+- **From the two reviews of the c11 chain** (0620–0627): the Bonsai 2 CUDA pool is released on every exit of the graph
+  evaluate; the legacy Prism Q2_0 layout is named also when its tensors come last; the MTP cost curve relearns after a
+  sustained rise; after a runtime head drop no reader reports a drafter; the measure pass skips every backend's host tail.
+- **A server with `--sleep-idle-seconds` on a GPU wakes** (0619, in c11 since r3).
+- **Interface**: the `ggml` digest of this cut is new (the two Cactus types in 0616, the library entries of 0645): every GPU
+  library and the Metal library of this index are new builds; a c11 library is refused by the engine with the digest named.
+  The `media` and `loader` digests are unchanged: the media sidecars and the macOS loader are the c11 files.
+
+Gated on these bytes (2026-10-13): bs2 (RTX PRO 6000 Blackwell, sm_120) Bonsai 2 CUDA + Vulkan gate 56 / 56 — KLD against the reference CUDA base 0.000168 (CUDA) / 0.000311 (Vulkan), needle 3 / 3 at 112k, PQ2_0 tg 127.0 / PTQ1_0 140.4 tok/s; the glibc 2.28 image, Vulkan and CUDA; Windows 11 RTX 3090 13 / 13 and Whistle 48 / 48; Mac mini Whistle 254 / 254 (Metal + CPU, word error rate 17.41 % against the reference engine's 17.65 % on 827 words, 16.08 % with the detector), signal stop 9 / 9, single-prompt teardown 7 / 7.
+
+Known at this cut: Whistle's live words on a loaded host swing between runs (the final text does not); the Realtime
+session on Windows with a GPU has not been run; chat mode on macOS was not run for the 0654 fix; no coopmat2 mat-mat for
+the ternary types on Vulkan and no Metal kernels for them (CPU fallback), as in c11.

components/engine/2610130818001/opencoti-0.10.5-c12-2610130818001

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f41db412a29b82690b77358e74c8ec000c76a202194c1c90aa8169765a7fcef9
+size 136540201

components/engine/2610130818001/opencoti-0.10.5-c12-2610130818001.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f41db412a29b82690b77358e74c8ec000c76a202194c1c90aa8169765a7fcef9
+size 136540201

components/macos/2610131352005/BUILD_INFO.macos.md

@@ -0,0 +1,41 @@
+# macos 2610131352005
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+Built from the engine source of build 2610130822001; gated floor (engine-min) 2610130818001; gated with 2610130822001.
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `ape-macos-aarch64` | macos-aarch64 | ape | `07f550415a59415973ef877eec99a776dfa5263b06fdb1fd9cefb7b884f0f723` | 56896 |
+| `ggml-metal-aarch64.dylib` | macos-aarch64 | metal | `54415d062a889c5cee87befb38cf41a754ac214ff3b3c9aedad87a0de67e58bb` | 1633248 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+ABI `loader` = `d6ff34022f1d3fb5af56ca0c80e3f4e552bf09f1bebdf0cb5f314a454bf38dbc` (exported) — sha256 of this list:
+
+```
+.cosmocc/4.0.2/bin/ape-m1.c 78af79e20abbb3f99a97355959b59bd58949564064adb754f0302ebabb4bf7a3
+```
+
+Apple notarisation (Accepted), submission id(s): `4a384d1f-2c01-4bbf-ac86-77651d119e7c eea7c66d-e8a4-4860-b9da-71e9416a2805`.
+
+The prebuilt APE loader (the c11 file, `loader` digest unchanged) and a NEW prebuilt Metal library, built on the Mac by the
+c12 engine from its embedded sources: it states the new `ggml` digest, declines the weight types it has no kernel for (0646)
+and computes the gated-delta-net forms it supports (0648). Notarised: Metal library
+4a384d1f-2c01-4bbf-ac86-77651d119e7c, loader eea7c66d-e8a4-4860-b9da-71e9416a2805 (Accepted).
+Gated on the Mac mini with the 2610130822001 engine and THIS signed library in the executable directory (`metal: loaded … (executable directory, 1633248 bytes)`): Whistle 254 / 254 (the encoder on MTL0), Qwen3.5-2B generation, signal stop 9 / 9, single-prompt teardown 7 / 7.

components/macos/2610131352005/ggml-metal-aarch64.dylib

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54415d062a889c5cee87befb38cf41a754ac214ff3b3c9aedad87a0de67e58bb
+size 1633248

components/media/2610071902001/BUILD_INFO.media.md

@@ -2,7 +2,7 @@
 
 Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
 
-Built from the engine source of build 2610071819001; gated floor (engine-min) 2610040950001; gated with 2610050707001, 2610071819001, 2610100921001, 2610101319001.
+Built from the engine source of build 2610071819001; gated floor (engine-min) 2610040950001; gated with 2610050707001, 2610071819001, 2610100921001, 2610100940001, 2610101328001.
 
 | file | platform | kind | sha256 | bytes |
 |---|---|---|---|---|
@@ -32,6 +32,13 @@ llamafile/oc-codec/oc_codec_sidecar.h f2c9001d1414aa2a55533b867b79c52a52c682558e
 
 Apple notarisation (Accepted), submission id(s): `83ee464c-3816-4b5c-9da7-cd4313840f25 aa53f9d9-14ba-443a-a059-64f82416f86a`.
 
+The twelve sidecars of c11, unchanged (same `media` digest): oc-codec, oc-audiocpp, oc-espeak, three per platform.
+Not re-gated as a set at this cut (same bytes, same digest). xollama loaded the codec sidecar beside the c12 engine on Linux x86_64 (mp3 input to Whistle, 2026-10-13).
+
+---
+
+The c11 notes:
+
 The twelve sidecars of c10, unchanged (same `media` digest): oc-codec, oc-audiocpp (0007), oc-espeak, three per platform.
 
 ---

components/sbsa/2610131352003/BUILD_INFO.sbsa.md

@@ -0,0 +1,29 @@
+# sbsa 2610131352003
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+Built from the engine source of build 2610130822001; gated floor (engine-min) 2610130818001; gated with nothing — NOT run.
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `ggml-cuda-sbsa-aarch64.so` | aarch64 | cuda | `7ba0366f6ca6bf42526e4f3032d6cb17eeb2ad1319df2960ab91f768f12963ce` | 346870848 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+CUDA 13 for Linux aarch64 (SBSA: DGX Spark / GB10 = sm_121, Jetson Thor). Cold build at chain 0652.
+**Never run: there is no such machine in the test fleet.**

components/sbsa/2610131352003/ggml-cuda-sbsa-aarch64.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ba0366f6ca6bf42526e4f3032d6cb17eeb2ad1319df2960ab91f768f12963ce
+size 346870848

components/vulkan/2610131352004/BUILD_INFO.vulkan.md

@@ -0,0 +1,31 @@
+# vulkan 2610131352004
+
+Component of the opencoti pin format v2 (docs/protocols/PIN_FORMAT.md).
+
+Built from the engine source of build 2610130822001; gated floor (engine-min) 2610130818001; gated with 2610130818001, 2610130822001.
+
+| file | platform | kind | sha256 | bytes |
+|---|---|---|---|---|
+| `ggml-vulkan-x86_64.so` | x86_64 | vulkan | `30d02746c876e119a2122b60f46c4fd02c2ac86832e248f7f3fa3fdf9bec3175` | 61300504 |
+| `ggml-vulkan-win-x86_64.dll` | win-x86_64 | vulkan | `16daf81ab4a685bd1b82a2b3712e8b5005e5ea40129ee26e63cb0b6367b791d6` | 62339097 |
+
+ABI `ggml` = `bac5050ddc8c58c51f0b059a7f9753b33af0ed79d355090e7c9a17ffaaaf340a` (exported) — sha256 of this list:
+
+```
+llama.cpp/ggml/include/ggml-alloc.h 94e4cd069b9313b2ceb35dacec901981e0bb478d8bb31035b7126be091998c23
+llama.cpp/ggml/include/ggml-backend.h 764d1640ad62579513c4837539708f89e8603c9dbeae053a13d1cbd84a82c865
+llama.cpp/ggml/include/ggml-cuda.h 674ea064021cf5b9970e78f1209e513dd23350b94a44a705a27eec1423fe4b0a
+llama.cpp/ggml/include/ggml-metal.h 322f36cd30f3e9e7aad7b5b9bc63078012fd0b7706ac4e24f5721b604f3d8980
+llama.cpp/ggml/include/ggml-neo-pipeline.h 190f39e183542d16588a3104d8cb4359bf52fc7a62afe8d8cc4c822f4085086d
+llama.cpp/ggml/include/ggml-oc-kvarn-range.h 76635fdd8978632f670640ddf5125a666441e3d47c680bb8b5c0a099e4861202
+llama.cpp/ggml/include/ggml-oc-ledger.h c1be11d69a57a1e721f1ae8720b87887d239ac0183d4277fbaa07f1be257d20c
+llama.cpp/ggml/include/ggml-vulkan.h f9df4f7161a84ce7600d53c3fac8a62837d5716676c9f38cd340699ef1c070d9
+llama.cpp/ggml/include/ggml.h ecfbf114f9e9151d22e30ef6d5139570c5c41b595843ca4f99125bd0196709aa
+llama.cpp/ggml/include/opencoti-devtype.h 94dea946173449bb76bffd7f733ab822fe658fe7198992de750ac012d29308ec
+llama.cpp/ggml/src/ggml-backend-impl.h f9fdfe6bae61d1387593b40fc0f90ceee490c6a339ec6d7388c51e4fca318d8f
+llamafile/gpu_backend.h 4929c36c034b6589e88653bcb7a5c5a225743a07320f64c0857d92bdaf048208
+```
+
+Built from the clean-room tree of chain 0652; the Linux library is byte-identical to the one every c12 dev build since 0645
+ran with. NEW `ggml` digest. Carries the refused-device lines of `--list-devices` (0637) and the logger level (0645).
+Gated with the 2610130822001 engine: the Vulkan cases of the bs2 56 / 56 gate (KLD 0.000311 against the CUDA base, needle 3 / 3 at 112k, tg 81.7 / 83.7 tok/s; two non-Bonsai models BIT-IDENTICAL between the c11 library and this one); the glibc 2.28 image; Windows RTX 3090 Vulkan (41.7 tok/s, the adapter names the Vulkan device, no library line without -lv). AMD Radeon on Windows and the refused-device lines on Windows: not run by us at this cut.

components/vulkan/2610131352004/ggml-vulkan-win-x86_64.dll

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16daf81ab4a685bd1b82a2b3712e8b5005e5ea40129ee26e63cb0b6367b791d6
+size 62339097

components/vulkan/2610131352004/ggml-vulkan-x86_64.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30d02746c876e119a2122b60f46c4fd02c2ac86832e248f7f3fa3fdf9bec3175
+size 61300504